The environmental cost of AI and machine learning
Machine learning methods have taken tremendous steps in the last decade. State of the at is truly impressive. At the same time, I feel empty inside. Somehow the methodology feels hollow. Take more data, more training and a more complex network. Voilà, you get better results. Hurray.
I think that the missing dimension is efficiency. Throwing more data or more computing power at the problem is an inefficient solution. It is not a scientific solution to a problem. Science is about understanding, not about brute-force.
To make machine learning more scientific, I therefore propose that we should use resource consumption as a measure of performance. Within resources, I would include at least computing power and amount of data. In addition to making machine learning more scientific, optimizing resource consumption would naturally make the field also environmentally more reasonable.